Environment Setup
conda create -n dust3r python=3.11 cmake=3.14.0 conda activate dust3r conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia # use the correct version of cuda for your system pip install -r requirements.txt # Optional: you can also install additional packages to: # - add support for HEIC images pip install -r requirements_optional.txt # DUST3R relies on RoPE positional embeddings for which you can compile some cuda kernels for faster runtime. HIGHLY RECOMMENDED. cd dust3r/croco/models/curope/ python setup.py build_ext --inplace cd ../../../
Dataset Download
Download the dataset and create the train/test split using this command:
cd datasets && python3 setup_12scenes.py
Model download
I downloaded the pretrained model using following command. You can also use models uploaded on huggingface.
wget https://download.europe.naverlabs.com/ComputerVision/DUSt3R/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth -P naver/
Knowledge Distillation Training
Start knowledge distillation training using the following command:
python3 knowledge-distillation/training.py --weights_path naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth --dataset_path datasets --scene_type 12scenes_apt1_kitchen >> logfile.log